However, in the recently released TensorFlow 1.4, we found that this module has been integrated into the core module, indicating that the interface has stabilized and is trending towards wider adoption. Therefore, I rewrote the program from section 4.4 using Estimator, keeping the variable names mostly unchanged. The code is as follows:

import tensorflow as tf
from sklearn import datasets, metrics, preprocessing
import numpy as np
import pandas as pd
import os

df = pd.read_csv("data/CHD.csv", header=0)
print( df.describe())

X = df['age'].astype(float)

feature_columns = [tf.contrib.layers.real_valued_column("X", dimension=1)]

classifier = tf.estimator.LinearClassifier(feature_columns=feature_columns,
                                    model_dir=os.path.join(".","tmp","logistic"))

input_fn_train = tf.estimator.inputs.numpy_input_fn(
                x={"X" : np.array(X)},
                y=np.array(df['chd']),
                batch_size=2,
                num_epochs=None,
                shuffle=True)

classifier.train(input_fn=input_fn_train, steps=2000)

Model Accuracy

score = classifier.evaluate(input_fn=input_fn_train, steps=50)["accuracy"]
print("Accuracy: %f" % score)

Note: This program can run on Ubuntu and MacOS, but not on Windows due to path issues. This seems to be a bug with Estimator, as it also doesn’t work under contrib.learn.

The most challenging part to write is the input_fn function, which is also the most important function. In this program, I directly used numpy_input_fn to construct it. Reference [1] provides another method from pandas that you can try yourself.

The advantage of input_fn is that it allows data to be read in a producer-consumer model. For a detailed explanation, you can refer to [2]. In simple terms, since IO is generally slow, we need to read data during the data processing phase to save time efficiently. This design uses multithreading to continuously fetch data in the background.

Building feature_columns requires some skill, mainly referenced from [3].

Another change is that model accuracy is no longer measured using the metric module, but rather the built-in module of Estimator.

If anyone has any questions, feel free to leave a message.

Practical Machine Learning Projects with TensorFlow

This article is sourced from the Asynchronous Community. Read the original text for more details, and the copyright belongs to the author.